A Collaborative Anticancer Drug Prediction Method Based on Knowledge Graph Embedding, GNN, and Bi-LSTM

Abstract

This invention provides a collaborative prediction method for anticancer drugs based on knowledge graph embedding, GNN, and Bi-LSTM, belonging to the field of collaborative prediction technology for anticancer drugs. To address the problem of poor collaborative prediction performance, the technical solution is as follows: Corresponding relationships between drugs, proteins, cell lines, and tissues are obtained from a database; a knowledge graph is constructed based on these relationships, and all drug combinations are grouped according to their collaborative scores; based on the drug combination grouping results and the constructed knowledge graph, a set of neighboring entities that interact with the drugs or cell lines is extracted, and an attention layer is constructed in the knowledge graph to iteratively update the initial entity embedding representation of each entity in the neighboring entity set; the entity neighborhood structure generated by the iterative update is used as input, and a graph neural network is used to extract higher-order structures and semantic relationships to learn the drug and its neighborhood representation; this invention is applied to collaborative prediction of anticancer drugs.

Description

Technical Field

[0001] This invention provides a collaborative prediction method for anticancer drugs based on knowledge graph embedding, GNN, and Bi-LSTM, belonging to the field of collaborative prediction technology for anticancer drugs. Background Technology

[0002] With the development of bioinformatics and deep learning technologies, the research and development of anticancer drugs has entered a new stage. Currently, traditional anticancer drug development methods rely on a large number of laboratory studies and clinical trials, which are time-consuming, costly, and have a relatively low success rate. According to research, in order to improve the efficiency and success rate of anticancer drug development, multiple drugs can be used in combination to reduce drug toxicity. This method is an effective anticancer drug development strategy.

[0003] However, there are still some limitations in predicting the effects of drug combination therapy. For example, in anticancer drug research, traditional machine learning methods often ignore the complexity of drug action and the network characteristics of biological systems. This limits the model's in-depth understanding of the drug synergy mechanism and results in a large prediction error for the synergistic effect of drugs. Summary of the Invention

[0004] In order to overcome the shortcomings of the prior art, the technical problem to be solved by the present invention is to provide a method for synergistic prediction of anticancer drugs based on knowledge graph embedding, GNN and Bi-LSTM.

[0005] To address the aforementioned technical problems, the present invention employs the following technical solution: a collaborative prediction method for anticancer drugs based on knowledge graph embedding, GNN, and Bi-LSTM, comprising the following prediction steps:

[0006] Step S1: Obtain from the database the corresponding relationships of drug-protein interactions, cell line-protein interactions, protein-protein interactions, and cell line-tissue interactions.

[0007] Step S2: Construct a knowledge graph based on the above correspondence, and group all drug combinations based on the synergistic scores of different drug combinations;

[0008] Step S3: Based on the drug combination grouping results and the constructed knowledge graph, extract the set of neighbor entities that interact with the drug or cell line, construct the attention layer of the knowledge graph, and iteratively update the initial entity embedding representation of each entity in the neighbor entity set.

[0009] Step S4: Using the entity neighborhood structure generated by the iterative update as input, extract the higher-order structure and semantic relationships through a graph neural network to learn the representation of drugs and their neighborhoods;

[0010] Step S5: Construct an attention-based pooling module to learn the interaction information between drug pairs. By strengthening the representation of drug pairs, the embedded representations are then aggregated into a single vector.

[0011] Step S6: Use a Bi-LSTM network to perform regression analysis on drug and cell embeddings, and then connect the analysis to a multilayer fully connected network to predict the collaborative score;

[0012] Step S7: Calculate the error using the MSELoss function, train the model multiple times using the Adam optimizer, and output the final anti-cancer drug combination scheme.

[0013] The specific method for constructing the knowledge graph and grouping according to the collaborative score in step S2 is as follows:

[0014] Step S21: Screen for groups of drug combinations that have potential predictive associations with cell lines;

[0015] Step S22: Each sample consists of two anticancer drugs, one cancer cell, and a related synergy score. The synergy score in DrugCombDB is defined as the ZIP value, and the synergy score in Oncology-Screen is defined as the Loewe additive value.

[0016] Group the drugs according to their scores, and then sort all drug combinations in descending order of their synergy scores:

[0017] The drug combinations corresponding to the first quarter of the synergistic score are defined as synergistic effects, and the drug combinations corresponding to the second quarter of the synergistic score are defined as antagonistic effects.

[0018] Step S23: The knowledge graph displays information about the association between drugs and proteins, cell lines and proteins, cell lines and tissues, and proteins and proteins. When information is passed to the target protein and cell line through the knowledge graph, the embedding of the target protein and cell line will increase the neighbor information of other relationships.

[0019] Step S24: Represent the knowledge graph as a triple T = (h, t, r) of multiple entities and relations. The specific method is as follows:

[0020] Define a knowledge graph of drug i-drug j-cell line, expressed as:

[0021] G = (ε, R);

[0022] Where ε represents drugs, proteins, cell lines, and tissues, and R represents the four types of corresponding relationships obtained in step S1;

[0023] Define h, t∈ε in the triple as the head entity and the tail entity, respectively;

[0024] In a triple, r∈R represents the relationship between two entities.

[0025] The specific method for iteratively updating the initial entity embedding representation of each entity in the neighbor entity set in step S3 is as follows:

[0026] Step S31: Find the entity set related to the drug / cell line in the knowledge graph and define the drug / cell line as parameter m;

[0027] Step S32: Recursively represent the set of adjacent entities of m as:

[0028]

[0029] Where l is the propagation depth of KG, (h,t,r) represents a set of triples, G is the knowledge graph, and L is the total propagation depth of the knowledge graph;

[0030] Step S33: Construct entity adjacency matrices A based on the entity sets. e Adjacency matrix A r , of which (A e A r ∈R v*n ), where v refers to the number of entities and n refers to the fixed number of neighboring entities;

[0031] In the adjacency matrix A r This includes the relationship between each entity and its neighboring entities, which includes the four types of corresponding relationships obtained in step S1;

[0032] Step S34: Based on the grouping results of the knowledge graph and drug combination, use the propagation mechanism of the knowledge graph to collect high-order association information with drugs / cell lines, and improve the vector representation of entities.

[0033] The specific method for extracting high-order structures and semantic relationships through graph neural networks and learning drug and its neighborhood representations in step S4 is as follows:

[0034] Step S41: Learn the embedding representation of the nodes, introduce a graph neural network, and perform neighborhood sampling for each node;

[0035] Step S42: Learn neighborhood information and set parameter H, which is defined as the receptive field in CNN;

[0036] Setting the receptive field depth of a given drug pair to 2 indicates that the case of nodes with second-order connectivity is considered.

[0037] Step S43: Obtain the final embedded representation of the entity by aggregating neighborhood information, and define the nodes directly connected to the drug node as N. neigh (e) uses a fixed neighborhood range S(e).

[0038] The specific method for constructing and aggregating the attention-based pooling module to learn the interaction information between drug pairs in step S5 is as follows:

[0039] Step S51: Input the aggregated entities into the constructed attention-based pooling module;

[0040] Step S52: The attention-based pooling module assigns a score to the substructure of each drug and performs a weighted summation of the embeddings of all nodes to obtain a graph-level representation;

[0041] Step S53: After sampling, the entity's own embedding representation and the embedding representation of its neighborhood information are aggregated using three aggregation methods to finally obtain the embedding representation of the current entity:

[0042] The final entity representation is aggregated using three aggregation methods, with the output being a single vector. The `sum` aggregation method is defined as an overlay operation, the `concat` aggregation method as a concatenation operation, and the `neighbor` aggregation method considers only neighborhood information while ignoring its own information. The corresponding aggregation expressions are as follows:

[0043] Sum aggregation:

[0044] concat aggregation:

[0045] neighbor aggregation:

[0046] Where σ is the Sigmoid activation function; W and b are trainable weights and biases; e is a certain entity, and S(e) is the neighborhood range.

[0047] The specific method for predicting the collaborative score in step S6 is as follows:

[0048] Specifically, a bidirectional long short-term memory network (Bi-LSTM) is used to predict the collaborative score:

[0049] Step S61: Input sequence preparation, the input sequence is divided into time steps of equal length, and each time step is a feature vector;

[0050] Step S62: Construct a Bi-LSTM model, including a forward LSTM and a backward LSTM. The forward LSTM is responsible for processing from the beginning of the sequence backward, and the backward LSTM is responsible for processing from the end of the sequence forward. Each LSTM unit includes an input gate, a forget gate, and an output gate component.

[0051] Step S63: Forward propagation, input the input sequence into the forward LSTM step by step to obtain the forward hidden state sequence for each time step;

[0052] Step S64: Backpropagation, input the input sequence into the backward LSTM step by step to obtain the backward hidden state sequence for each time step;

[0053] Step S65: Merge hidden states. Merge the forward and backward hidden state sequences according to the time step. The merged hidden state sequence contains the comprehensive information at each time step in the sequence, and also considers the information from both the beginning and end of the sequence.

[0054] Step S66: Output layer processing, inputting the merged hidden state sequence into two fully connected layers.

[0055] The specific method for the model to continuously learn and output the scheme in step S7 is as follows:

[0056] Specifically, MSELoss is used to calculate the error, and the Adam optimizer is used for training:

[0057] Step S71: Calculate the difference between the predicted value and the true value. The difference for each sample is a vector or matrix.

[0058] Step S72: Calculate the squared difference by squaring each difference value to obtain the corresponding squared error;

[0059] Step S73: Calculate the mean by summing the squared errors of all samples and then dividing by the number of samples to obtain the average squared error, which is the value of the mean squared error loss function.

[0060] Step S74: Backpropagation. Using the mean squared error as the loss, calculate the gradient of the loss function with respect to the model parameters through the backpropagation algorithm.

[0061] Step S75: Update the model parameters using the Adam optimizer based on the gradient of the loss function, making the model approximate the optimal solution:

[0062] Based on the gradient of the loss function, the learning rate lr, the momentum parameter β1, and the exponential decay rate β2 used to calculate the second moment of the gradient, the first moment estimate is updated first, then the second moment estimate is updated, and then the bias between the first and second moment estimates is corrected, the parameters are updated, and the iteration is repeated.

[0063] The advantages of this invention compared to existing technologies are as follows: This invention proposes a collaborative prediction method for anticancer drugs based on knowledge graph embedding, graph neural network training, and bidirectional long short-term memory network prediction. The knowledge graph used is a heterogeneous graph, where nodes represent entities and edges represent relationships between entities. The rich semantic connections between entities can help the system discover potential relationships between entities. The knowledge graph is used to learn entities such as drugs, proteins, and their topological neighbors to obtain more stable embedding representations. A graph neural network is used to receive the topological information of each entity in the knowledge graph and learn features through an attention-based pooling module to better understand module information. A Bi-LSTM model is used to capture the interaction between drug combinations and cells. Two inverse LSTM layers are intricately arranged together, and their bidirectional structure allows the model to consider past and future information simultaneously. After regression analysis, the output is projected into a fully connected network to obtain a collaborative score, thereby improving the prediction accuracy and efficiency of anticancer drug combinations and optimizing the screening and evaluation process of drug combinations in clinical applications. Attached Figure Description

[0064] The present invention will be further described below with reference to the accompanying drawings:

[0065] Figure 1 This is a flowchart of the steps in the anticancer drug synergistic prediction method of the present invention;

[0066] Figure 2 This is a diagram illustrating the model structure of the neural network used in the prediction method of this invention. Detailed Implementation

[0067] like Figure 1 and Figure 2 As shown, this invention proposes a collaborative prediction method for anticancer drug combinations based on knowledge graph embedding, graph neural network training (GNN), and bidirectional long short-term memory network (Bi-LSTM). This prediction method can effectively improve the prediction accuracy and efficiency of anticancer drug combinations and optimize the screening and evaluation process of drug combinations in clinical applications. Specifically, this method constructs a knowledge graph, uses graph neural network training, learns interaction information based on attention-based pooling modules, performs regression analysis using Bi-LSTM networks, and finally outputs a collaborative score through a fully connected network, thereby achieving intelligent and efficient prediction of the effects of anticancer drug combinations. The method specifically includes the following steps:

[0068] Step S1: Collect and obtain four types of entities and their corresponding relationships;

[0069] Among them, the four types of entities include drugs, proteins, cell lines, and tissues;

[0070] Correspondence refers to: protein-drug-protein interactions, cell line-protein interactions, protein-protein interactions, and cell line-tissue interactions.

[0071] The data used in this invention comes from publicly available databases, specifically:

[0072] Drug-protein interactions were collected from six commonly used databases: DrugBank, BindingDB, and ChEMBL, which contained 15,015 drug-protein associations.

[0073] Cell line-protein interactions were obtained from response data in the Cancer Cell Line Encyclopedia (CCLE), which included 18,022 protein-coding genes, 1,035 cancer cell lines, and 749,551 cell line-protein associations.

[0074] Protein-protein interactions are derived from the human protein-protein interaction genome.

[0075] Cell line-tissue interactions are derived from DrugCombDB and Oncology-Screen.

[0076] Step S2: Preprocess the data and group it according to the collaboration score;

[0077] First, groups with potential predictive associations between drug combinations and cell lines were screened. Each sample consisted of two anticancer drugs, one cancer cell, and related synergy scores. In DrugCombDB, the synergy score was the ZIP value, and in Oncology-Screen, the synergy score was the Loewe additive value.

[0078] All drug combinations are grouped based on their synergy scores: all drug combinations are sorted in descending order of their synergy scores.

[0079] The drug combinations corresponding to the first quarter of the synergy score are classified as synergistic, while the drug combinations corresponding to the second quarter of the synergy score are classified as antagonistic. This can exclude low-confidence combination drug samples.

[0080] Synergistic effects were considered positive samples, and antagonistic effects were considered negative samples. The preprocessed data were then divided into training and testing sets in an 8:2 ratio to train and evaluate the effectiveness of the treatment recommendation model.

[0081] Then construct the knowledge graph:

[0082] The knowledge graph displays information about drug-protein, cell line-protein, cell line-tissue, and protein-protein relationships. When information is passed to a target protein and cell line through this knowledge graph, their embeddings add information about their neighbors and other relationships.

[0083] A knowledge graph can be represented as multiple entity-relationship-entity triples T = (h, r, T). Specifically, the knowledge graph of drug i-drug j-cell line can be represented as G = (ε, R), where ε refers to the four types of entities (drug, protein, cell line, tissue), and R refers to the four types of relationships (drug-protein, cell line-protein, cell line-tissue, and protein-protein relationship). In the triple, h, T ∈ ε represent the head entity and the tail entity, respectively, and r ∈ R represents the relationship between the two entities.

[0084] Step S3: Extract the neighbor entity set and refine the initial entity embedding representation:

[0085] Neighboring entities in a knowledge graph are always highly correlated. This module is set up to obtain multi-hop neighbor nodes for drugs / cell lines.

[0086] Step S31: First, find the entity set related to the drug / cell line in the knowledge graph. In this embodiment of the invention, the drug / cell line is abbreviated as m. Here, the adjacent entities of m can be recursively represented as:

[0087]

[0088] Where l is the propagation depth of KG, (h,t,r) represents a set of triples, where G is the knowledge graph and L is the total propagation depth of the knowledge graph;

[0089] Step S33: Subsequently, construct the entity adjacency matrix A based on the entity set. e Adjacency matrix A r , of which (A e A r ∈R v*n In the adjacency matrix, v represents the number of entities, and n represents a fixed number of neighboring entities. The relational adjacency matrix includes the relationships between each entity and its neighboring entities, categorized into four types: drug-protein, protein-protein, cell line-protein, and cell line-tissue.

[0090] Step S34: Finally, based on the grouping results of the knowledge graph and drug combination, the propagation mechanism of the knowledge graph is used to collect high-order association information with drugs / cell lines, thereby improving and enhancing the vector representation of entities.

[0091] Step S4: Build a network model containing a graph neural network (GNN) and an attention-based pooling module to encode drug features and neighborhood structures between drug pairs. The graph neural network extracts higher-order structures and semantic relationships, learning representations of drugs and their neighborhoods.

[0092] Step S41: Directly learning the embedding representation of nodes does not consider the rich neighborhood information of each entity. A graph neural network is introduced to sample the neighborhood of each node.

[0093] Step S42: In the process of learning neighborhood information, in order to learn more neighborhood information, considering the neighborhood range of each drug node's two hops, a parameter H is set, which can be understood as the receptive field in CNN. In this invention, the receptive field depth of a given drug pair is set to 2, indicating that the case of second-order connected nodes is considered. Of course, H can take a larger value.

[0094] Step S43: Obtain the final embedded representation of the entity by aggregating neighborhood information, and define the nodes directly connected to the drug node as N. neigh (e). Since the neighborhood distribution of each drug node is different, a fixed neighborhood range S(e) is adopted for computational convenience by using the GraphSAGE computation method.

[0095] Step S5: Construct an attention-based pooling module to learn and aggregate the interaction information between drug pairs:

[0096] Step S51: Input the aggregated entities into the constructed attention-based pooling module;

[0097] Step S52: The aggregated entities described above are fed into the pooling module as input. Attention-based pooling helps the model consider which substructures within the chemical substances are more important for predicting synergistic drugs. By using the attention-based pooling module, not only can interaction information between drug pairs be obtained, but also important chemical substructures of the drugs can be identified. The attention-based pooling module assigns a score to each drug substructure and performs a weighted summation of the embeddings of all nodes to obtain a graph-level representation, expressed as:

[0098]

[0099] In the above formula, and These are the graph embedding matrices of drug x and drug x, respectively, taken from the last layer of the GNN, and obtained by the tanh() function to obtain A. x A y , representing the fraction assigned to each substructure of the drug. And a x and ay This represents the attention score for each drug pair, obtained by summing each score and then applying softmax() activation. Here, M represents the node count in the graph. The final graph-level representation is obtained by weighting the vectors of all nodes based on their attention scores, expressed as:

[0100] g x =multiply(a x H x );

[0101] g y =multiply(a y H y );

[0102] Therefore, for any entity node, its representation can be enhanced through the above step S52, expressed as:

[0103] e=attention-basedpooling(E);

[0104] The attention-based pooling() method represents the pooling operation described above.

[0105] Step S53: During the aggregation process, the entity's own embedding representation and the embedding representation of its neighborhood information are aggregated using three aggregation methods to finally obtain the embedding representation of the current entity.

[0106] The final entity representation is aggregated using three aggregation methods, outputting a single vector. The `sum` aggregation method is a superposition operation, `concat` is a concatenation operation, and `neighbor` only considers neighborhood information while ignoring its own information. The expression is as follows:

[0107] Sum aggregation:

[0108] concat aggregation:

[0109] neighbor aggregation:

[0110] Where σ is the Sigmoid activation function; W and b are trainable weights and biases; e is a certain entity, and S(e) is the neighborhood range.

[0111] Step S6: Use a bidirectional long short-term memory network (Bi-LSTM) for prediction. Specific steps include:

[0112] Step S61: Input sequence preparation. Divide the input sequence into time steps of equal length. Typically, each time step is a feature vector.

[0113] Step S62: Construct a Bi-LSTM model. The forward LSTM is responsible for processing the sequence from the beginning to the end, while the backward LSTM is responsible for processing the sequence from the end to the beginning. Each LSTM unit includes components such as an input gate, a forget gate, and an output gate.

[0114] Step S63: Forward propagation. Input the input sequence into the forward LSTM step by step to obtain the forward hidden state sequence at each time step. The forward hidden state sequence represents the state information at each time step in the sequence and can capture temporal dependencies;

[0115] Step S64: Backpropagation. Input the input sequence into the backward LSTM step by step to obtain the backward hidden state sequence at each time step. The backward hidden state sequence reflects the state information at each time step in the sequence.

[0116] Step S65: Merge the hidden states. Merge the forward and backward hidden state sequences by time step, usually by simply concatenating them position by position. The merged hidden state sequence contains comprehensive information at each time step in the sequence, taking into account information from both the beginning and end of the sequence.

[0117] Step S66: Output layer processing, inputting the merged hidden state sequence into two fully connected layers;

[0118] Since the effects of drug-cell interactions are not affected by the order of administration, this setting is beneficial for the task, where the future context of the sequence is just as important as the past context.

[0119] Step S7: Calculate the error using MSELoss and train using the Adam optimizer. Specific steps include:

[0120] Step S71: Calculate the difference between the predicted value and the true value. For each sample, this difference can be a vector or a matrix.

[0121] Step S72: Calculate the squared difference by squaring each difference value to obtain the corresponding squared error. The purpose is to eliminate positive and negative differences and focus only on the magnitude of the difference between the predicted and the true values.

[0122] Step S73: Calculate the mean by summing the squared errors of all samples and then dividing by the number of samples to obtain the average squared error. This is the value of the mean squared error loss function.

[0123] Step S74: Backpropagation. Using the mean squared error as the loss, the gradient of the loss function with respect to the model parameters is calculated through the backpropagation algorithm. This updates the model parameters, minimizing the loss function.

[0124] Step S75: Use the Adam optimizer to update the model parameters based on the gradient of the loss function so that the model approaches the optimal solution;

[0125] The model is trained using the Adam optimizer with a maximum of 100,000 iterations. Based on the gradient of the loss function, the learning rate (lr), the momentum parameter (β1), the exponential decay rate (β2) used to calculate the second moment of the gradient, and the parameter stability parameter (ε), the first-order moment estimate is updated first, followed by the second-order moment estimate. Then, the bias between the first and second-order moment estimates is corrected, the parameters are updated, and the iteration is repeated. The Adam optimizer adaptively adjusts the learning rate based on the historical gradients of each parameter; different parameters have different learning rates, thus enabling more efficient learning.

[0126] Through multiple rounds of training and learning, the model is continuously optimized and relevant weight parameters are adjusted. This model effectively improves the prediction accuracy and efficiency of anticancer drug combinations, thereby optimizing the screening and evaluation process of drug combinations in clinical applications. Ultimately, it can provide patients with treatment suggestions tailored to their individual circumstances, including measures such as changing drugs or adjusting dosages for different drug combinations, to assist doctors in providing more personalized and efficient treatment.

[0127] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for anticancer drug synergy prediction based on knowledge graph embedding, GNN and Bi-LSTM, characterized in that: The prediction steps include the following: Step S1: Obtain the corresponding relationships of drug-protein interaction, cell line-protein interaction, protein-protein interaction and cell line-tissue interaction from the database respectively; Step S2: Construct a knowledge graph based on the above-mentioned corresponding relationships, and group all drug combinations based on the synergy scores of different drug combinations; Step S3: Based on the grouping results of drug combinations and the constructed knowledge graph, extract the neighbor entity set that interacts with the drug or cell line, construct the attention layer of the knowledge graph, and iteratively update the initial entity embedding representation of each entity in the neighbor entity set; Step S4: Take the entity neighborhood structure generated by the iterative update as input, extract high-order structure and semantic relationship through the graph neural network, and learn the drug and its neighborhood representation, specifically: Step S41: Learn the embedding representation of the node, introduce the graph neural network, and sample the neighborhood of each node; Step S42: Learn the neighborhood information, set the parameter H, and define it as the perception domain in CNN; Set the acceptance field depth of a given drug pair to 2, which means considering the case of second-order connected nodes; Step S43: Obtain the final embedding representation of the entity by aggregating neighborhood information, define the nodes directly connected to the drug node as N neigh (e), using a fixed neighborhood range S(e); Step S5: Construct a pooling module based on attention to learn the interaction information between drug pairs, and aggregate the embedding representation into a single vector by strengthening the representation of drug pairs, specifically: Step S51: Input the aggregated entity into the constructed pooling module based on attention; Step S52: The pooling module based on attention assigns a score to each drug substructure, and performs weighted summation on the embedding of all nodes to obtain the graph-level representation; Step S53: After sampling, the embedding representation of the entity itself and the embedding representation of the neighborhood information are aggregated through three aggregation methods to obtain the final embedding representation of the entity: The final entity representation is aggregated through three aggregation methods, and the output is a single vector, where the sum aggregation method is a superposition operation, the concat aggregation method is a splicing operation, and the neighbor aggregation method only considers neighborhood information and ignores its own information. The corresponding aggregation expression is: sum aggregation: ; concat aggregation: ; neighbor aggregation: ; Where σ is the Sigmod activation function; W and b are trainable weights and biases; e is a certain entity, and S(e) is the neighborhood range; Step S6: Perform regression analysis on the drug and cell embedding using a Bi-LSTM network, and then connect it to a multi-layer fully connected network to predict the synergy score, specifically: Specifically, use a bidirectional long short-term memory network (Bi-LSTM) to predict the synergy score: Step S61: Input sequence preparation, divide the input sequence into equal-length time steps, and each time step is a feature vector; Step S62: Construct a Bi-LSTM model, including forward LSTM and backward LSTM, where forward LSTM is responsible for processing from the start of the sequence to the end, and backward LSTM is responsible for processing from the end of the sequence to the beginning. Each LSTM unit includes an input gate, a forget gate, and an output gate component; Step S63: forward propagation, input the input sequence into the forward LSTM one time step at a time, and obtain the forward hidden state sequence of each time step; Step S64: backward propagation, input the input sequence into the backward LSTM one time step at a time, and obtain the backward hidden state sequence of each time step; Step S65: merge the hidden states, merge the forward and backward hidden state sequences by time step, and the merged hidden state sequence contains the comprehensive information of each time step in the sequence, considering the information from both the beginning and the end of the sequence; Step S66: output layer processing, input the merged hidden state sequence into two fully connected layers; Step S7: calculate the error using the MSELoss function, and use the Adam optimizer to train the model for multiple rounds, and output the final anticancer drug combination scheme, the specific method is: Specifically, use MSELoss to calculate the error and use Adam optimizer to train: Step S71: calculate the difference between the predicted value and the true value, and the difference of each sample is a vector or matrix; Step S72: calculate the squared difference, take the square of each difference value to obtain the corresponding squared error; Step S73: calculate the mean, sum all the squared errors of the samples, and then divide by the number of samples to obtain the average squared error, which is the value of the mean square error loss function; Step S74: back propagation, use the mean square error as the loss, and calculate the gradient of the loss function with respect to the model parameters through the back propagation algorithm; Step S75: use the Adam optimizer to update the model parameters according to the gradient of the loss function, so that the model approaches the optimal solution: According to the gradient of the loss function, the learning rate lr, the momentum parameter β1, the exponential decay rate β2 used to calculate the second moment of the gradient, update the first moment estimate, then update the second moment estimate, and then correct the bias of the first and second moment estimates, update the parameters, and repeat the iteration.

2. The knowledge graph embedding, GNN and Bi-LSTM based anticancer drug synergistic prediction method according to claim 1, characterized in that: The specific method of constructing the knowledge graph and grouping according to the synergy score in step S2 is: Step S21: filter out the drug combination and cell line with potential predictive relevance; Step S22: each sample consists of two anticancer drugs, a cancer cell and the related synergy score, define the synergy score in DrugCombDB as ZIP value, and the synergy score in Oncology-Screen as Loewe additive value; Group according to the size relationship of the score, arrange all drug combinations in descending order of synergy score: Define the first quarter of the synergy score as synergistic effect, and the last quarter of the synergy score as antagonistic effect; Step S23: the knowledge graph displays the information of drug and protein, cell line and protein, cell line and tissue, and protein and protein association, when an information is transmitted to the target protein and cell line through the knowledge graph, the embedding of the target protein and cell line will increase the neighbor information of other relationships; Step S24: express the knowledge graph as a triple T=(h,t,r) of multiple entities and relationships, the specific method is: Define the knowledge graph of drug i-drug j-cell line, the expression is: G = (ε, R) ; Where ε is a drug, protein, cell line, tissue, R is the four types of corresponding contact obtained in step S1; Define h, t ∈ ε in the triple, which represent the head entity and tail entity respectively; Define r ∈ R in the triple, which represents the relationship between the two entities.

3. The anti-cancer drug synergistic prediction method based on knowledge graph embedding, GNN and Bi-LSTM according to claim 2, characterized in that: The specific method of iterative updating of the initial entity embedding representation of each entity in the neighbor entity set in step S3 is: Step S31: Find the entity set related to the drug / cell line in the knowledge graph, and define the drug / cell line as the parameter m; Step S32: Recursively represent the neighbor entity set of m as: ； Where l is the propagation depth of KG, (h, t, r) represents a set of triples, G is the knowledge graph, and L is the total propagation depth of the knowledge graph; Step S33: Constructing entity adjacency matrix A according to the entity set respectively e and relationship adjacency matrix A r , where (A e , A r ∈R v*n ), v refers to the number of entities, and n refers to the number of fixed neighbor entities; In the relationship adjacency matrix A r includes the relationship of each entity with its neighbor entity, the relationship includes the four types of corresponding contacts obtained in step S1; Step S34: Based on the grouping results of the knowledge graph and the drug combination, use the propagation mechanism of the knowledge graph to collect high-order association information related to the drug / cell line, and improve the vector representation of the entity.

Images